Synopsis

Keywords: Parallel Transmit & Multiband, Parallel Transmit & Multiband

Ultra-high field MRI with parallel transmit offers significant advantages but B₁⁺ maps are required to utilize its potential. Acquiring B₁⁺ maps for each channel is time-consuming and limits the application of ultra-high field in clinical practice. In this abstract we present an approach to acquire full 3D multi-channel B₁⁺ maps in 11 seconds. We do this by employing a time-interleaved acquisition of modes strategy using undersampled relative maps and two fast undersampled Sandwich B₁⁺ maps all reconstructed using TxLR, a calibrationless reconstruction for transmit field maps. We demonstrate this approach with simulations and scans in phantom and in the brain.

Introduction

B₁⁺ inhomogeneity is severe across the heart and large across the brain at 7T, limiting the utility of ultrahigh-field cardiac imaging. Parallel transmit (pTx) solves this problem but requires multi-channel B₁⁺ maps which are difficult to acquire in a clinically reasonable timeframe and are compounded by issues of cardiorespiratory motion. An adaptation to the presaturation TurboFLASH method¹ of acquiring B₁⁺ maps, termed Sandwich, enables faster absolute maps² that are less sensitive to T₁ and maintain flow-insensitivity. In this work we combine three acquisition strategies, Sandwich, B1TIAMO³, and relative transmit maps and reconstruct the data using transmit low rank reconstruction (TxLR)⁴ to measure multi-channel B₁⁺ maps. Our is aim to reduce the acquisition duration to a single breath-hold or 16 heartbeats. We evaluate the method using simulations, phantom and in vivo brain data.

Methods

The proposed acquisition scheme is shown in Figure 1. This includes three measurement sections: 1) an undersampled 3D coil-cycled multi-channel relative transmit map, acquired using a spoiled low flip angle gradient echo sequence, one image from each transmit channel is acquired in an interleaved fashion, minimising the influence of magnetisation history. 2) and 3) are 3D undersampled absolute maps with complementary shim modes (CP⁺ and CP²⁺).

Simulations were performed using retrospectively undersampled in-silico data to establish the achievable acceleration and masking strategy. The in-silico data were generated in Sim4Life (3.4, ZMT) using a model of an 8-channel dipole array around DUKES heart, channel-independent complex Gaussian noise was added to this data. K-space was cropped to 24×24 phase encodes and randomly undersampled using a Poisson disk distribution at rates of 1-16. The central 4×4 region of k-space was maintained.

A 7 Tesla MRI (VB17, Siemens Healthcare) equipped with parallel transmit was used for data acquisition. Data were measured in a water phantom and in vivo brain of a healthy volunteer using an 8Tx/32Rx head coil (Nova Medical).

The 3D relative maps were acquired with TE_GRE=1.31ms, TR_GRE=3.2ms, bandwidth=490Hz/px, FOV=(250mm)³, matrix=24×24×24, resolution=(10.4mm)³, RF-spoiling, non-selective RF excitation FA=9°, and a total acceleration factor 6 with central 4×4 central region and a different undersampling mask for each transmit channel.

The 3D absolute maps were acquired using the Sandwich² scheme based on a TurboFLASH sequence, TE=1.26ms, TR=3ms, bandwidth=490Hz/px, FOV=(250mm)³, matrix=24×24×24, resolution=(10.4mm)³, RF-spoiling and total acceleration factor 6 (4×4 central calibration region) with non-selective RF excitation (β). A non-adiabatic hyperbolic secant RF pulse was used as our presaturation (α) with nominal α/β flip angles 90/4.5°, each volume was acquired in 3 segments, each with 32 lines. This was repeated for two acquisition modes, with the first mode preceded by two dummy TRs. The total acquisition time was 8s (2 dummy TRs + 3 segments mode 1 + 3 segments mode 2). Different undersampling masks were used for each shim mode, but the reference and prepared images were acquired using the same mask, with centre-out encoding. The total acquisition time to acquire both the relative and absolute maps was 11 seconds.

The relative, reference and prepared cartesian undersampled k-spaces were concatenated to form a 24×24×24×8×12 (N_kx×N_ky×N_kz×N_Rx×N_Tx) matrix and jointly reconstructed using the TxLR method⁴ sliced along the readout direction. The k-space data were Hann filtered, zero-padded to 24×32×32 and fast Fourier transformed back to full image space. The receive channel images were combined using sensitivity maps estimated using ESPIRiT⁵. Reconstruction was performed offline in MATLAB (R2021a, MathWorks) on an Intel Xeon 28×2.40GHz 128Gb RAM computer. We compare a slice from our 3D maps to a 2D presaturation TurboFLASH method which took 90 seconds to acquire.

Results

Simulation results from synthetic 8Tx/8Rx body images in Figure 2 show acceleration factors up to 6 are achievable with a mean difference to the fully sampled of <1°. Figure 3a shows the 3D dataset acquired in a phantom in 11 seconds. Figure 3b and 3c compares our 3D data to a fully sampled 2D dataset with the difference in part d. Figure 4 shows similar data in the brain of one subject with a mean absolute difference of 0.4°. Reconstruction took 12 minutes.

Discussion

Simulations show that acceleration factors of 6 are comfortably achievable and higher are possible. We found that using a small fully sampled region (4×4) aided the reconstruction allowing faster reconstruction times and reduced RMSE, particularly for higher acceleration factors, this also improved our in vivo results. The accelerated data appears spatially smoother than the fully sampled images, indicating further investigation of actual resolution is required. The main limitation currently is the 12 minute reconstruction time for the 3D datasets on our hardware. Although we do not show any in vivo cardiac data, our simulations and brain work suggests that to acquire our accelerated 3D maps in the heart would 8 heartbeats for relative mapping and 8 for absolute mapping, even taking into account the need for segmenting and gating the relative maps. In principle, the two sequences can be acquired together in a single breath-hold or in two subsequent breath-holds. Going forward we aim to apply this method in the heart.

Conclusion

We obtain 3D multi-channel absolute B₁⁺ maps in the brain in 11 seconds for an entire transmit array with good accuracy to fully sampled reference maps.

Acknowledgements

JK acknowledges the support of EPSRC through an iCASE award in collaboration with Siemens Healthcare Ltd. The Wellcome Centre for Integrative Neuroimaging is supported by core funding from the Wellcome Trust (203139/Z/16/Z). LV is funded by a Sir Henry Dale Fellowship awarded jointly by the Wellcome Trust and the Royal Society (221805/Z/20/Z) and also acknowledges the support of the Slovak Grant Agencies VEGA (2/0003/20) and APVV (#19–0032). MC is supported by the Canada Research Chairs Program.